Species distribution models (SDMs) are considered valuable risk assessment tools in invasive species management. However, early detection and containment are crucial management strategies, and few studies have evaluated the impact of temporal sampling bias on the utility of SDMs created early in a biological invasion. While the abundance and geographic range of occurrence records may increase with time after invasion, resource managers often cannot wait for a destructive species to establish before modeling its potential extent. We assessed the relative predictive ability of models trained on chronologically cumulative sets of early occurrence records for invasive shot-hole borers (Euwallacea spp., Coleoptera: Curculionidae: Scolytinae), first detected in southern California in 2010. Using presence records collected from 2012 to 2016, we developed a series of SDMs for ISHB using Maxent, selected for its ability to produce reliable models with relatively few survey records. Twenty models were created using five cumulative sampling periods, with and without spatial sample bias correction, and two spatial extents: 1. California and 2. Ecoregions 8 and 85 (EPA Level III). Predictors included variables for topography, climate, and host tree characteristics.
Results/Conclusions
Predictive performance was strong among all SDMs, as indicated by high AUC values (AUC≥0.91). Spatial bias correction and spatial extent had little effect on AUC, emphasizing the need to assess presence-only SDMs with additional metrics. We used the maximum training sensitivity plus specificity threshold to compare model sensitivity (i.e. presence records correctly classified as true-positives). All models were able to predict over 89% of true-positives within their sampling period, indicating high sensitivity regardless of the degree of temporal sample bias. On average, 2014 and 2015 models were also able to predict 68% and 78% of true-positives from 2016, respectively, with higher sensitivity among models using the California extent. This suggests that SDMs for emerging invasive species can be useful in identifying areas vulnerable to invasion, provided that input data are continuously updated as new occurrence records become available. These models may also aid and focus monitoring and containment efforts during the early stages of biological invasion. Ongoing research will assess models using additional comparative metrics, including field validation. Where early survey data is available, future research may investigate the effect of temporal sampling bias on SDMs for other invasive species. This would help to establish a monitoring time threshold required to develop reliable SDMs for early detection.